disco icon indicating copy to clipboard operation
disco copied to clipboard

Add validation and data shuffle

Open Nacho114 opened this issue 2 years ago • 6 comments

Need to add data validation and shuffling to the data loader. (only inside each local client data, not between clients that is)

  • [x] Update the return type of load to Promise<Data>
  • [x] Add shuffle and seed mechanism (the TFJS version seemed to be broken, might be worth to test it again though)
  • [ ] Add validation dataset, there is a validationSplit number param, but it does not seem to be supported on the current TFJS version we have. So either update TFJS and try it out, or do a split manually in the data loader, and add corresponding dataSplit param in builder.

Nacho114 avatar Apr 27 '22 13:04 Nacho114

To avoid things breaking it's best to be conservative for now, and just add the easier solution for validation set, and we can in a later point try to update tfjs

Nacho114 avatar Apr 27 '22 13:04 Nacho114

data shuffling is only done locally (not between clients)

Nacho114 avatar Apr 27 '22 14:04 Nacho114

#338 indirectly depends on this

s314cy avatar Jun 30 '22 08:06 s314cy

did you guys check what's currently shown on the validation plot during training? if there is an issue we can remove it temporarily (discussed with hugo)

martinjaggi avatar Jun 30 '22 08:06 martinjaggi

#356 adds shuffling

s314cy avatar Jul 05 '22 13:07 s314cy

#372 adds validation split for image datasets

s314cy avatar Jul 13 '22 11:07 s314cy

done i guess in #356 and #499

martinjaggi avatar Oct 25 '22 15:10 martinjaggi